direct.c 25.3 KB
Newer Older
L
Linus Torvalds 已提交
1 2 3 4 5 6 7 8 9
/*
 * linux/fs/nfs/direct.c
 *
 * Copyright (C) 2003 by Chuck Lever <cel@netapp.com>
 *
 * High-performance uncached I/O for the Linux NFS client
 *
 * There are important applications whose performance or correctness
 * depends on uncached access to file data.  Database clusters
10
 * (multiple copies of the same instance running on separate hosts)
L
Linus Torvalds 已提交
11
 * implement their own cache coherency protocol that subsumes file
12 13 14
 * system cache protocols.  Applications that process datasets
 * considerably larger than the client's memory do not always benefit
 * from a local cache.  A streaming video server, for instance, has no
L
Linus Torvalds 已提交
15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36
 * need to cache the contents of a file.
 *
 * When an application requests uncached I/O, all read and write requests
 * are made directly to the server; data stored or fetched via these
 * requests is not cached in the Linux page cache.  The client does not
 * correct unaligned requests from applications.  All requested bytes are
 * held on permanent storage before a direct write system call returns to
 * an application.
 *
 * Solaris implements an uncached I/O facility called directio() that
 * is used for backups and sequential I/O to very large files.  Solaris
 * also supports uncaching whole NFS partitions with "-o forcedirectio,"
 * an undocumented mount option.
 *
 * Designed by Jeff Kimmel, Chuck Lever, and Trond Myklebust, with
 * help from Andrew Morton.
 *
 * 18 Dec 2001	Initial implementation for 2.4  --cel
 * 08 Jul 2002	Version for 2.4.19, with bug fixes --trondmy
 * 08 Jun 2003	Port to 2.5 APIs  --cel
 * 31 Mar 2004	Handle direct I/O without VFS support  --cel
 * 15 Sep 2004	Parallel async reads  --cel
37
 * 04 May 2005	support O_DIRECT with aio  --cel
L
Linus Torvalds 已提交
38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55
 *
 */

#include <linux/errno.h>
#include <linux/sched.h>
#include <linux/kernel.h>
#include <linux/file.h>
#include <linux/pagemap.h>
#include <linux/kref.h>

#include <linux/nfs_fs.h>
#include <linux/nfs_page.h>
#include <linux/sunrpc/clnt.h>

#include <asm/system.h>
#include <asm/uaccess.h>
#include <asm/atomic.h>

56
#include "internal.h"
C
Chuck Lever 已提交
57 58
#include "iostat.h"

L
Linus Torvalds 已提交
59 60
#define NFSDBG_FACILITY		NFSDBG_VFS

61
static struct kmem_cache *nfs_direct_cachep;
L
Linus Torvalds 已提交
62 63 64 65 66 67

/*
 * This represents a set of asynchronous requests that we're waiting on
 */
struct nfs_direct_req {
	struct kref		kref;		/* release manager */
68 69

	/* I/O parameters */
70
	struct nfs_open_context	*ctx;		/* file open context info */
71
	struct kiocb *		iocb;		/* controlling i/o request */
72
	struct inode *		inode;		/* target file of i/o */
73 74

	/* completion state */
75
	atomic_t		io_count;	/* i/os we're waiting for */
76 77
	spinlock_t		lock;		/* protect completion state */
	ssize_t			count,		/* bytes actually processed */
L
Linus Torvalds 已提交
78
				error;		/* any reported error */
79
	struct completion	completion;	/* wait for i/o completion */
80 81

	/* commit state */
82
	struct list_head	rewrite_list;	/* saved nfs_write_data structs */
83 84 85 86 87
	struct nfs_write_data *	commit_data;	/* special write_data for commits */
	int			flags;
#define NFS_ODIRECT_DO_COMMIT		(1)	/* an unstable reply was received */
#define NFS_ODIRECT_RESCHED_WRITES	(2)	/* write verification failed */
	struct nfs_writeverf	verf;		/* unstable write verifier */
L
Linus Torvalds 已提交
88 89
};

90
static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode);
91 92 93 94 95 96 97 98 99 100 101 102
static const struct rpc_call_ops nfs_write_direct_ops;

static inline void get_dreq(struct nfs_direct_req *dreq)
{
	atomic_inc(&dreq->io_count);
}

static inline int put_dreq(struct nfs_direct_req *dreq)
{
	return atomic_dec_and_test(&dreq->io_count);
}

L
Linus Torvalds 已提交
103
/**
104 105 106 107 108 109 110 111 112 113 114
 * nfs_direct_IO - NFS address space operation for direct I/O
 * @rw: direction (read or write)
 * @iocb: target I/O control block
 * @iov: array of vectors that define I/O buffer
 * @pos: offset in file to begin the operation
 * @nr_segs: size of iovec array
 *
 * The presence of this routine in the address space ops vector means
 * the NFS client supports direct I/O.  However, we shunt off direct
 * read and write requests before the VFS gets them, so this method
 * should never be called.
L
Linus Torvalds 已提交
115
 */
116 117 118
ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, loff_t pos, unsigned long nr_segs)
{
	dprintk("NFS: nfs_direct_IO (%s) off/no(%Ld/%lu) EINVAL\n",
119
			iocb->ki_filp->f_path.dentry->d_name.name,
120
			(long long) pos, nr_segs);
121 122 123 124

	return -EINVAL;
}

125
static void nfs_direct_dirty_pages(struct page **pages, unsigned int pgbase, size_t count)
126
{
127
	unsigned int npages;
128
	unsigned int i;
129 130 131 132 133

	if (count == 0)
		return;
	pages += (pgbase >> PAGE_SHIFT);
	npages = (count + (pgbase & ~PAGE_MASK) + PAGE_SIZE - 1) >> PAGE_SHIFT;
134 135
	for (i = 0; i < npages; i++) {
		struct page *page = pages[i];
136
		if (!PageCompound(page))
137
			set_page_dirty(page);
138
	}
139 140
}

141
static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
142
{
143
	unsigned int i;
144 145
	for (i = 0; i < npages; i++)
		page_cache_release(pages[i]);
146 147
}

148
static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
L
Linus Torvalds 已提交
149
{
150 151
	struct nfs_direct_req *dreq;

152
	dreq = kmem_cache_alloc(nfs_direct_cachep, GFP_KERNEL);
153 154 155 156
	if (!dreq)
		return NULL;

	kref_init(&dreq->kref);
157
	kref_get(&dreq->kref);
158
	init_completion(&dreq->completion);
159
	INIT_LIST_HEAD(&dreq->rewrite_list);
160
	dreq->iocb = NULL;
161
	dreq->ctx = NULL;
162
	spin_lock_init(&dreq->lock);
163
	atomic_set(&dreq->io_count, 0);
164 165
	dreq->count = 0;
	dreq->error = 0;
166
	dreq->flags = 0;
167 168

	return dreq;
L
Linus Torvalds 已提交
169 170
}

171
static void nfs_direct_req_free(struct kref *kref)
L
Linus Torvalds 已提交
172 173
{
	struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
174 175 176

	if (dreq->ctx != NULL)
		put_nfs_open_context(dreq->ctx);
L
Linus Torvalds 已提交
177 178 179
	kmem_cache_free(nfs_direct_cachep, dreq);
}

180 181 182 183 184
static void nfs_direct_req_release(struct nfs_direct_req *dreq)
{
	kref_put(&dreq->kref, nfs_direct_req_free);
}

185 186 187 188 189
/*
 * Collects and returns the final error value/byte-count.
 */
static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
{
190
	ssize_t result = -EIOCBQUEUED;
191 192 193 194 195

	/* Async requests don't wait here */
	if (dreq->iocb)
		goto out;

196
	result = wait_for_completion_interruptible(&dreq->completion);
197 198

	if (!result)
199
		result = dreq->error;
200
	if (!result)
201
		result = dreq->count;
202 203 204 205 206

out:
	return (ssize_t) result;
}

207
/*
208 209
 * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
 * the iocb is still valid here if this is a synchronous request.
210 211 212 213
 */
static void nfs_direct_complete(struct nfs_direct_req *dreq)
{
	if (dreq->iocb) {
214
		long res = (long) dreq->error;
215
		if (!res)
216
			res = (long) dreq->count;
217
		aio_complete(dreq->iocb, res, 0);
218 219
	}
	complete_all(&dreq->completion);
220

221
	nfs_direct_req_release(dreq);
222 223
}

224
/*
225 226 227
 * We must hold a reference to all the pages in this direct read request
 * until the RPCs complete.  This could be long *after* we are woken up in
 * nfs_direct_wait (for instance, if someone hits ^C on a slow server).
228
 */
T
Trond Myklebust 已提交
229
static void nfs_direct_read_result(struct rpc_task *task, void *calldata)
L
Linus Torvalds 已提交
230
{
T
Trond Myklebust 已提交
231
	struct nfs_read_data *data = calldata;
L
Linus Torvalds 已提交
232 233
	struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;

T
Trond Myklebust 已提交
234 235
	if (nfs_readpage_result(task, data) != 0)
		return;
236 237

	spin_lock(&dreq->lock);
238
	if (unlikely(task->tk_status < 0)) {
239
		dreq->error = task->tk_status;
240 241 242 243 244 245 246 247 248
		spin_unlock(&dreq->lock);
	} else {
		dreq->count += data->res.count;
		spin_unlock(&dreq->lock);
		nfs_direct_dirty_pages(data->pagevec,
				data->args.pgbase,
				data->res.count);
	}
	nfs_direct_release_pages(data->pagevec, data->npages);
249 250 251

	if (put_dreq(dreq))
		nfs_direct_complete(dreq);
L
Linus Torvalds 已提交
252 253
}

T
Trond Myklebust 已提交
254 255 256 257 258
static const struct rpc_call_ops nfs_read_direct_ops = {
	.rpc_call_done = nfs_direct_read_result,
	.rpc_release = nfs_readdata_release,
};

259
/*
260 261 262 263 264
 * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
 * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
 * bail and stop sending more reads.  Read length accounting is
 * handled automatically by nfs_direct_read_result().  Otherwise, if
 * no requests have been sent, just return an error.
L
Linus Torvalds 已提交
265
 */
266 267 268
static ssize_t nfs_direct_read_schedule_segment(struct nfs_direct_req *dreq,
						const struct iovec *iov,
						loff_t pos)
L
Linus Torvalds 已提交
269
{
270
	struct nfs_open_context *ctx = dreq->ctx;
271
	struct inode *inode = ctx->path.dentry->d_inode;
272 273
	unsigned long user_addr = (unsigned long)iov->iov_base;
	size_t count = iov->iov_len;
274
	size_t rsize = NFS_SERVER(inode)->rsize;
275 276 277 278
	unsigned int pgbase;
	int result;
	ssize_t started = 0;

L
Linus Torvalds 已提交
279
	do {
280
		struct nfs_read_data *data;
281
		size_t bytes;
L
Linus Torvalds 已提交
282

283 284 285
		pgbase = user_addr & ~PAGE_MASK;
		bytes = min(rsize,count);

286
		result = -ENOMEM;
287
		data = nfs_readdata_alloc(nfs_page_array_len(pgbase, bytes));
288 289 290 291 292 293 294
		if (unlikely(!data))
			break;

		down_read(&current->mm->mmap_sem);
		result = get_user_pages(current, current->mm, user_addr,
					data->npages, 1, 0, data->pagevec, NULL);
		up_read(&current->mm->mmap_sem);
295 296 297 298 299
		if (result < 0) {
			nfs_readdata_release(data);
			break;
		}
		if ((unsigned)result < data->npages) {
300 301 302 303 304 305 306 307
			bytes = result * PAGE_SIZE;
			if (bytes <= pgbase) {
				nfs_direct_release_pages(data->pagevec, result);
				nfs_readdata_release(data);
				break;
			}
			bytes -= pgbase;
			data->npages = result;
308 309 310
		}

		get_dreq(dreq);
311

312
		data->req = (struct nfs_page *) dreq;
L
Linus Torvalds 已提交
313 314 315 316
		data->inode = inode;
		data->cred = ctx->cred;
		data->args.fh = NFS_FH(inode);
		data->args.context = ctx;
317
		data->args.offset = pos;
L
Linus Torvalds 已提交
318
		data->args.pgbase = pgbase;
319
		data->args.pages = data->pagevec;
L
Linus Torvalds 已提交
320 321 322 323 324
		data->args.count = bytes;
		data->res.fattr = &data->fattr;
		data->res.eof = 0;
		data->res.count = bytes;

T
Trond Myklebust 已提交
325 326
		rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
				&nfs_read_direct_ops, data);
L
Linus Torvalds 已提交
327 328 329 330 331 332
		NFS_PROTO(inode)->read_setup(data);

		data->task.tk_cookie = (unsigned long) inode;

		rpc_execute(&data->task);

C
Chuck Lever 已提交
333 334
		dprintk("NFS: %5u initiated direct read call "
			"(req %s/%Ld, %zu bytes @ offset %Lu)\n",
L
Linus Torvalds 已提交
335 336 337 338 339 340
				data->task.tk_pid,
				inode->i_sb->s_id,
				(long long)NFS_FILEID(inode),
				bytes,
				(unsigned long long)data->args.offset);

341 342
		started += bytes;
		user_addr += bytes;
343
		pos += bytes;
344
		/* FIXME: Remove this unnecessary math from final patch */
L
Linus Torvalds 已提交
345 346
		pgbase += bytes;
		pgbase &= ~PAGE_MASK;
347
		BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
L
Linus Torvalds 已提交
348 349 350

		count -= bytes;
	} while (count != 0);
351 352

	if (started)
353
		return started;
354
	return result < 0 ? (ssize_t) result : -EFAULT;
L
Linus Torvalds 已提交
355 356
}

357 358 359 360 361 362 363 364 365 366 367 368 369
static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
					      const struct iovec *iov,
					      unsigned long nr_segs,
					      loff_t pos)
{
	ssize_t result = -EINVAL;
	size_t requested_bytes = 0;
	unsigned long seg;

	get_dreq(dreq);

	for (seg = 0; seg < nr_segs; seg++) {
		const struct iovec *vec = &iov[seg];
370
		result = nfs_direct_read_schedule_segment(dreq, vec, pos);
371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389
		if (result < 0)
			break;
		requested_bytes += result;
		if ((size_t)result < vec->iov_len)
			break;
		pos += vec->iov_len;
	}

	if (put_dreq(dreq))
		nfs_direct_complete(dreq);

	if (requested_bytes != 0)
		return 0;

	if (result < 0)
		return result;
	return -EIO;
}

390 391
static ssize_t nfs_direct_read(struct kiocb *iocb, const struct iovec *iov,
			       unsigned long nr_segs, loff_t pos)
L
Linus Torvalds 已提交
392
{
393
	ssize_t result = 0;
L
Linus Torvalds 已提交
394
	sigset_t oldset;
395
	struct inode *inode = iocb->ki_filp->f_mapping->host;
L
Linus Torvalds 已提交
396 397 398
	struct rpc_clnt *clnt = NFS_CLIENT(inode);
	struct nfs_direct_req *dreq;

399
	dreq = nfs_direct_req_alloc();
L
Linus Torvalds 已提交
400 401 402
	if (!dreq)
		return -ENOMEM;

C
Chuck Lever 已提交
403
	dreq->inode = inode;
404
	dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
405 406
	if (!is_sync_kiocb(iocb))
		dreq->iocb = iocb;
L
Linus Torvalds 已提交
407 408

	rpc_clnt_sigmask(clnt, &oldset);
409
	result = nfs_direct_read_schedule_iovec(dreq, iov, nr_segs, pos);
410 411
	if (!result)
		result = nfs_direct_wait(dreq);
L
Linus Torvalds 已提交
412
	rpc_clnt_sigunmask(clnt, &oldset);
413
	nfs_direct_req_release(dreq);
L
Linus Torvalds 已提交
414 415 416 417

	return result;
}

418
static void nfs_direct_free_writedata(struct nfs_direct_req *dreq)
L
Linus Torvalds 已提交
419
{
420 421
	while (!list_empty(&dreq->rewrite_list)) {
		struct nfs_write_data *data = list_entry(dreq->rewrite_list.next, struct nfs_write_data, pages);
422
		list_del(&data->pages);
423
		nfs_direct_release_pages(data->pagevec, data->npages);
424 425 426
		nfs_writedata_release(data);
	}
}
L
Linus Torvalds 已提交
427

428 429 430
#if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
{
431 432 433
	struct inode *inode = dreq->inode;
	struct list_head *p;
	struct nfs_write_data *data;
L
Linus Torvalds 已提交
434

435
	dreq->count = 0;
436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472
	get_dreq(dreq);

	list_for_each(p, &dreq->rewrite_list) {
		data = list_entry(p, struct nfs_write_data, pages);

		get_dreq(dreq);

		/*
		 * Reset data->res.
		 */
		nfs_fattr_init(&data->fattr);
		data->res.count = data->args.count;
		memset(&data->verf, 0, sizeof(data->verf));

		/*
		 * Reuse data->task; data->args should not have changed
		 * since the original request was sent.
		 */
		rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
				&nfs_write_direct_ops, data);
		NFS_PROTO(inode)->write_setup(data, FLUSH_STABLE);

		data->task.tk_priority = RPC_PRIORITY_NORMAL;
		data->task.tk_cookie = (unsigned long) inode;

		/*
		 * We're called via an RPC callback, so BKL is already held.
		 */
		rpc_execute(&data->task);

		dprintk("NFS: %5u rescheduled direct write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
				data->task.tk_pid,
				inode->i_sb->s_id,
				(long long)NFS_FILEID(inode),
				data->args.count,
				(unsigned long long)data->args.offset);
	}
473

474 475
	if (put_dreq(dreq))
		nfs_direct_write_complete(dreq, inode);
476 477 478 479 480 481 482 483 484 485 486
}

static void nfs_direct_commit_result(struct rpc_task *task, void *calldata)
{
	struct nfs_write_data *data = calldata;
	struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;

	/* Call the NFS version-specific code */
	if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
		return;
	if (unlikely(task->tk_status < 0)) {
487 488
		dprintk("NFS: %5u commit failed with error %d.\n",
				task->tk_pid, task->tk_status);
489
		dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
490
	} else if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
491 492
		dprintk("NFS: %5u commit verify failed\n", task->tk_pid);
		dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
L
Linus Torvalds 已提交
493 494
	}

495 496
	dprintk("NFS: %5u commit returned %d\n", task->tk_pid, task->tk_status);
	nfs_direct_write_complete(dreq, data->inode);
L
Linus Torvalds 已提交
497 498
}

499 500 501 502 503 504
static const struct rpc_call_ops nfs_commit_direct_ops = {
	.rpc_call_done = nfs_direct_commit_result,
	.rpc_release = nfs_commit_release,
};

static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
L
Linus Torvalds 已提交
505
{
506
	struct nfs_write_data *data = dreq->commit_data;
L
Linus Torvalds 已提交
507

508
	data->inode = dreq->inode;
509
	data->cred = dreq->ctx->cred;
L
Linus Torvalds 已提交
510

511
	data->args.fh = NFS_FH(data->inode);
512 513
	data->args.offset = 0;
	data->args.count = 0;
514 515 516
	data->res.count = 0;
	data->res.fattr = &data->fattr;
	data->res.verf = &data->verf;
L
Linus Torvalds 已提交
517

518 519 520
	rpc_init_task(&data->task, NFS_CLIENT(dreq->inode), RPC_TASK_ASYNC,
				&nfs_commit_direct_ops, data);
	NFS_PROTO(data->inode)->commit_setup(data, 0);
L
Linus Torvalds 已提交
521

522 523 524 525
	data->task.tk_priority = RPC_PRIORITY_NORMAL;
	data->task.tk_cookie = (unsigned long)data->inode;
	/* Note: task.tk_ops->rpc_release will free dreq->commit_data */
	dreq->commit_data = NULL;
L
Linus Torvalds 已提交
526

527
	dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
L
Linus Torvalds 已提交
528

529 530
	rpc_execute(&data->task);
}
L
Linus Torvalds 已提交
531

532 533 534
static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
{
	int flags = dreq->flags;
L
Linus Torvalds 已提交
535

536 537 538 539
	dreq->flags = 0;
	switch (flags) {
		case NFS_ODIRECT_DO_COMMIT:
			nfs_direct_commit_schedule(dreq);
L
Linus Torvalds 已提交
540
			break;
541 542 543 544 545 546 547
		case NFS_ODIRECT_RESCHED_WRITES:
			nfs_direct_write_reschedule(dreq);
			break;
		default:
			if (dreq->commit_data != NULL)
				nfs_commit_free(dreq->commit_data);
			nfs_direct_free_writedata(dreq);
548
			nfs_zap_mapping(inode, inode->i_mapping);
549 550 551
			nfs_direct_complete(dreq);
	}
}
L
Linus Torvalds 已提交
552

553 554
static void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
{
555
	dreq->commit_data = nfs_commit_alloc();
556 557 558 559 560 561 562 563
	if (dreq->commit_data != NULL)
		dreq->commit_data->req = (struct nfs_page *) dreq;
}
#else
static inline void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
{
	dreq->commit_data = NULL;
}
L
Linus Torvalds 已提交
564

565 566 567
static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
{
	nfs_direct_free_writedata(dreq);
568
	nfs_zap_mapping(inode, inode->i_mapping);
569 570 571
	nfs_direct_complete(dreq);
}
#endif
L
Linus Torvalds 已提交
572

573
static void nfs_direct_write_result(struct rpc_task *task, void *calldata)
L
Linus Torvalds 已提交
574
{
575 576 577 578 579 580 581
	struct nfs_write_data *data = calldata;
	struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
	int status = task->tk_status;

	if (nfs_writeback_done(task, data) != 0)
		return;

582
	spin_lock(&dreq->lock);
L
Linus Torvalds 已提交
583

584
	if (unlikely(status < 0)) {
585
		/* An error has occurred, so we should not commit */
586
		dreq->flags = 0;
587 588
		dreq->error = status;
	}
589 590
	if (unlikely(dreq->error != 0))
		goto out_unlock;
591 592

	dreq->count += data->res.count;
L
Linus Torvalds 已提交
593

594 595 596 597 598
	if (data->res.verf->committed != NFS_FILE_SYNC) {
		switch (dreq->flags) {
			case 0:
				memcpy(&dreq->verf, &data->verf, sizeof(dreq->verf));
				dreq->flags = NFS_ODIRECT_DO_COMMIT;
L
Linus Torvalds 已提交
599
				break;
600 601 602 603 604
			case NFS_ODIRECT_DO_COMMIT:
				if (memcmp(&dreq->verf, &data->verf, sizeof(dreq->verf))) {
					dprintk("NFS: %5u write verify failed\n", task->tk_pid);
					dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
				}
L
Linus Torvalds 已提交
605 606
		}
	}
607
out_unlock:
608
	spin_unlock(&dreq->lock);
L
Linus Torvalds 已提交
609 610
}

611 612 613
/*
 * NB: Return the value of the first error return code.  Subsequent
 *     errors after the first one are ignored.
L
Linus Torvalds 已提交
614
 */
615
static void nfs_direct_write_release(void *calldata)
L
Linus Torvalds 已提交
616
{
617 618
	struct nfs_write_data *data = calldata;
	struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
L
Linus Torvalds 已提交
619

620 621
	if (put_dreq(dreq))
		nfs_direct_write_complete(dreq, data->inode);
622 623 624 625
}

static const struct rpc_call_ops nfs_write_direct_ops = {
	.rpc_call_done = nfs_direct_write_result,
626
	.rpc_release = nfs_direct_write_release,
627 628 629
};

/*
630 631 632 633 634
 * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 * bail and stop sending more writes.  Write length accounting is
 * handled automatically by nfs_direct_write_result().  Otherwise, if
 * no requests have been sent, just return an error.
635
 */
636 637 638
static ssize_t nfs_direct_write_schedule_segment(struct nfs_direct_req *dreq,
						 const struct iovec *iov,
						 loff_t pos, int sync)
639
{
640
	struct nfs_open_context *ctx = dreq->ctx;
641
	struct inode *inode = ctx->path.dentry->d_inode;
642 643
	unsigned long user_addr = (unsigned long)iov->iov_base;
	size_t count = iov->iov_len;
644
	size_t wsize = NFS_SERVER(inode)->wsize;
645 646 647
	unsigned int pgbase;
	int result;
	ssize_t started = 0;
648

L
Linus Torvalds 已提交
649
	do {
650
		struct nfs_write_data *data;
651 652
		size_t bytes;

653 654 655
		pgbase = user_addr & ~PAGE_MASK;
		bytes = min(wsize,count);

656
		result = -ENOMEM;
657
		data = nfs_writedata_alloc(nfs_page_array_len(pgbase, bytes));
658 659 660 661 662 663 664
		if (unlikely(!data))
			break;

		down_read(&current->mm->mmap_sem);
		result = get_user_pages(current, current->mm, user_addr,
					data->npages, 0, 0, data->pagevec, NULL);
		up_read(&current->mm->mmap_sem);
665 666 667 668 669
		if (result < 0) {
			nfs_writedata_release(data);
			break;
		}
		if ((unsigned)result < data->npages) {
670 671 672 673 674 675 676 677
			bytes = result * PAGE_SIZE;
			if (bytes <= pgbase) {
				nfs_direct_release_pages(data->pagevec, result);
				nfs_writedata_release(data);
				break;
			}
			bytes -= pgbase;
			data->npages = result;
678 679 680 681
		}

		get_dreq(dreq);

682
		list_move_tail(&data->pages, &dreq->rewrite_list);
683

684
		data->req = (struct nfs_page *) dreq;
685 686 687 688
		data->inode = inode;
		data->cred = ctx->cred;
		data->args.fh = NFS_FH(inode);
		data->args.context = ctx;
689
		data->args.offset = pos;
690
		data->args.pgbase = pgbase;
691
		data->args.pages = data->pagevec;
692 693 694
		data->args.count = bytes;
		data->res.fattr = &data->fattr;
		data->res.count = bytes;
695
		data->res.verf = &data->verf;
696 697 698

		rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
				&nfs_write_direct_ops, data);
699
		NFS_PROTO(inode)->write_setup(data, sync);
L
Linus Torvalds 已提交
700

701 702
		data->task.tk_priority = RPC_PRIORITY_NORMAL;
		data->task.tk_cookie = (unsigned long) inode;
L
Linus Torvalds 已提交
703

704
		rpc_execute(&data->task);
L
Linus Torvalds 已提交
705

C
Chuck Lever 已提交
706 707
		dprintk("NFS: %5u initiated direct write call "
			"(req %s/%Ld, %zu bytes @ offset %Lu)\n",
708 709 710 711 712
				data->task.tk_pid,
				inode->i_sb->s_id,
				(long long)NFS_FILEID(inode),
				bytes,
				(unsigned long long)data->args.offset);
L
Linus Torvalds 已提交
713

714 715
		started += bytes;
		user_addr += bytes;
716
		pos += bytes;
717 718

		/* FIXME: Remove this useless math from the final patch */
719 720
		pgbase += bytes;
		pgbase &= ~PAGE_MASK;
721
		BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
L
Linus Torvalds 已提交
722

723 724
		count -= bytes;
	} while (count != 0);
725 726

	if (started)
727
		return started;
728
	return result < 0 ? (ssize_t) result : -EFAULT;
729
}
L
Linus Torvalds 已提交
730

731 732 733 734 735 736 737 738 739 740 741 742 743
static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
					       const struct iovec *iov,
					       unsigned long nr_segs,
					       loff_t pos, int sync)
{
	ssize_t result = 0;
	size_t requested_bytes = 0;
	unsigned long seg;

	get_dreq(dreq);

	for (seg = 0; seg < nr_segs; seg++) {
		const struct iovec *vec = &iov[seg];
744 745
		result = nfs_direct_write_schedule_segment(dreq, vec,
							   pos, sync);
746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764
		if (result < 0)
			break;
		requested_bytes += result;
		if ((size_t)result < vec->iov_len)
			break;
		pos += vec->iov_len;
	}

	if (put_dreq(dreq))
		nfs_direct_write_complete(dreq, dreq->inode);

	if (requested_bytes != 0)
		return 0;

	if (result < 0)
		return result;
	return -EIO;
}

765 766 767
static ssize_t nfs_direct_write(struct kiocb *iocb, const struct iovec *iov,
				unsigned long nr_segs, loff_t pos,
				size_t count)
768
{
769
	ssize_t result = 0;
770
	sigset_t oldset;
771
	struct inode *inode = iocb->ki_filp->f_mapping->host;
772 773
	struct rpc_clnt *clnt = NFS_CLIENT(inode);
	struct nfs_direct_req *dreq;
774 775
	size_t wsize = NFS_SERVER(inode)->wsize;
	int sync = 0;
L
Linus Torvalds 已提交
776

777
	dreq = nfs_direct_req_alloc();
778 779
	if (!dreq)
		return -ENOMEM;
780 781
	nfs_alloc_commit_data(dreq);

782 783
	if (dreq->commit_data == NULL || count < wsize)
		sync = FLUSH_STABLE;
L
Linus Torvalds 已提交
784

785
	dreq->inode = inode;
786
	dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
787 788
	if (!is_sync_kiocb(iocb))
		dreq->iocb = iocb;
L
Linus Torvalds 已提交
789

790
	rpc_clnt_sigmask(clnt, &oldset);
791
	result = nfs_direct_write_schedule_iovec(dreq, iov, nr_segs, pos, sync);
792 793
	if (!result)
		result = nfs_direct_wait(dreq);
794
	rpc_clnt_sigunmask(clnt, &oldset);
795
	nfs_direct_req_release(dreq);
L
Linus Torvalds 已提交
796 797 798 799 800 801 802

	return result;
}

/**
 * nfs_file_direct_read - file direct read operation for NFS files
 * @iocb: target I/O control block
803 804
 * @iov: vector of user buffers into which to read data
 * @nr_segs: size of iov vector
805
 * @pos: byte offset in file where reading starts
L
Linus Torvalds 已提交
806 807 808 809 810 811
 *
 * We use this function for direct reads instead of calling
 * generic_file_aio_read() in order to avoid gfar's check to see if
 * the request starts before the end of the file.  For that check
 * to work, we must generate a GETATTR before each direct read, and
 * even then there is a window between the GETATTR and the subsequent
812
 * READ where the file size could change.  Our preference is simply
L
Linus Torvalds 已提交
813 814
 * to do all reads the application wants, and the server will take
 * care of managing the end of file boundary.
815
 *
L
Linus Torvalds 已提交
816 817 818 819 820
 * This function also eliminates unnecessarily updating the file's
 * atime locally, as the NFS server sets the file's atime, and this
 * client must read the updated atime from the server back into its
 * cache.
 */
821 822
ssize_t nfs_file_direct_read(struct kiocb *iocb, const struct iovec *iov,
				unsigned long nr_segs, loff_t pos)
L
Linus Torvalds 已提交
823 824 825 826
{
	ssize_t retval = -EINVAL;
	struct file *file = iocb->ki_filp;
	struct address_space *mapping = file->f_mapping;
827 828 829 830
	size_t count;

	count = iov_length(iov, nr_segs);
	nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
L
Linus Torvalds 已提交
831

832
	dprintk("nfs: direct read(%s/%s, %zd@%Ld)\n",
833 834
		file->f_path.dentry->d_parent->d_name.name,
		file->f_path.dentry->d_name.name,
835
		count, (long long) pos);
L
Linus Torvalds 已提交
836 837 838 839 840

	retval = 0;
	if (!count)
		goto out;

T
Trond Myklebust 已提交
841 842 843
	retval = nfs_sync_mapping(mapping);
	if (retval)
		goto out;
L
Linus Torvalds 已提交
844

845
	retval = nfs_direct_read(iocb, iov, nr_segs, pos);
L
Linus Torvalds 已提交
846
	if (retval > 0)
847
		iocb->ki_pos = pos + retval;
L
Linus Torvalds 已提交
848 849 850 851 852 853 854 855

out:
	return retval;
}

/**
 * nfs_file_direct_write - file direct write operation for NFS files
 * @iocb: target I/O control block
856 857
 * @iov: vector of user buffers from which to write data
 * @nr_segs: size of iov vector
858
 * @pos: byte offset in file where writing starts
L
Linus Torvalds 已提交
859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877
 *
 * We use this function for direct writes instead of calling
 * generic_file_aio_write() in order to avoid taking the inode
 * semaphore and updating the i_size.  The NFS server will set
 * the new i_size and this client must read the updated size
 * back into its cache.  We let the server do generic write
 * parameter checking and report problems.
 *
 * We also avoid an unnecessary invocation of generic_osync_inode(),
 * as it is fairly meaningless to sync the metadata of an NFS file.
 *
 * We eliminate local atime updates, see direct read above.
 *
 * We avoid unnecessary page cache invalidations for normal cached
 * readers of this file.
 *
 * Note that O_APPEND is not supported for NFS direct writes, as there
 * is no atomic O_APPEND write facility in the NFS protocol.
 */
878 879
ssize_t nfs_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
				unsigned long nr_segs, loff_t pos)
L
Linus Torvalds 已提交
880
{
C
Chuck Lever 已提交
881
	ssize_t retval = -EINVAL;
L
Linus Torvalds 已提交
882 883
	struct file *file = iocb->ki_filp;
	struct address_space *mapping = file->f_mapping;
884
	size_t count;
L
Linus Torvalds 已提交
885

886 887 888 889
	count = iov_length(iov, nr_segs);
	nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);

	dfprintk(VFS, "nfs: direct write(%s/%s, %zd@%Ld)\n",
890 891
		file->f_path.dentry->d_parent->d_name.name,
		file->f_path.dentry->d_name.name,
892
		count, (long long) pos);
893

894 895
	retval = generic_write_checks(file, &pos, &count, 0);
	if (retval)
L
Linus Torvalds 已提交
896
		goto out;
897 898
	if (!count)
		goto out;	/* return 0 */
899 900 901

	retval = -EINVAL;
	if ((ssize_t) count < 0)
L
Linus Torvalds 已提交
902 903 904 905
		goto out;
	retval = 0;
	if (!count)
		goto out;
906

T
Trond Myklebust 已提交
907 908 909
	retval = nfs_sync_mapping(mapping);
	if (retval)
		goto out;
L
Linus Torvalds 已提交
910

911
	retval = nfs_direct_write(iocb, iov, nr_segs, pos, count);
912

L
Linus Torvalds 已提交
913
	if (retval > 0)
914
		iocb->ki_pos = pos + retval;
L
Linus Torvalds 已提交
915 916 917 918 919

out:
	return retval;
}

920 921 922 923
/**
 * nfs_init_directcache - create a slab cache for nfs_direct_req structures
 *
 */
D
David Howells 已提交
924
int __init nfs_init_directcache(void)
L
Linus Torvalds 已提交
925 926 927
{
	nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
						sizeof(struct nfs_direct_req),
928 929
						0, (SLAB_RECLAIM_ACCOUNT|
							SLAB_MEM_SPREAD),
930
						NULL);
L
Linus Torvalds 已提交
931 932 933 934 935 936
	if (nfs_direct_cachep == NULL)
		return -ENOMEM;

	return 0;
}

937
/**
D
David Howells 已提交
938
 * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
939 940
 *
 */
941
void nfs_destroy_directcache(void)
L
Linus Torvalds 已提交
942
{
943
	kmem_cache_destroy(nfs_direct_cachep);
L
Linus Torvalds 已提交
944
}